| Abstrakt: |
Early detection of accidents and rescue are of paramount importance in the reduction of fatalities. Social media data, which has evolved to become an important source of sharing information, plays a great role in building machine learning-based models for classifying posts related to accidents. Since the context of the word "accident" is difficult to determine in a posting, various works in literature have developed better classifiers for predicting whether the posting is actually related to an accident. However, an ensemble of classifiers are known to provide better performance than the basic models. Therefore, in this direction, we present a novel weighted majority voting-based ensemble approach for context classification of tweets (WM-ECCT) to detect whether the tweets are related or unrelated to road accidents. For the proposed ensemble model, the weighting scheme is based on the principle of false prediction to true prediction ratio. Also, the proposed model uses the multi-inducer technique and bootstrap sampling to reduce misclassification rates. Moreover, we propose a context-aware labeling approach for the annotation of tweets into related and unrelated categories. Experiments conducted reveal that the proposed ensemble model outperforms the different standalone machine learning and ensemble models on various performance measures. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink